An Explainable Multimodal Transformer Framework for Automated Fact-Checking and Fake News Detection using Large Language Models

This paper proposes an Explainable Multimodal Transformer Framework (EMTF) to automated fact checking, fake news; it aims to solve the increasing threat of multimodal misinformation with texts, visuals, and sounds. Current methods tend not to be interpretable, cross-domain or adaptive fused. The proposed EMTF will eliminate such constraints by combining pretrained modality-specific encoders and contrastive semantic alignment to achieve cross-modal consistency. A retrieval-augmented pipeline uses external evidence to verify grounded in a dynamic mechanism of adaptation of reliability-weighted fusion based on contributions of modalities, and a retrieval-augmented pipeline. Besides, a large language model (LLM) is used to produce prediction of veracity, as well as explanations which can be understood by humans. The framework is tested on benchmark databases of the Cartus Automated Fact-Checking Resources, FEVER, FakeSV, and r/Fakeddit. The results of experiments indicate that the accuracy, F1-score, and explainability are significantly better than recent baselines. Measures of explainability like sufficiency, comprehensiveness ensure that EMTF can generate faithful and reliable justifications. The suggested algorithm offers a system that is reproducible, scalable, and transparent in order to detect misinformation in a trustworthy manner.